PRIVACY IN VOIP NETWORKS: FLOW ANALYSIS ATTACKS AND DEFENSE

Page created by Bill Becker
 
CONTINUE READING
PRIVACY IN VOIP NETWORKS: FLOW ANALYSIS ATTACKS AND DEFENSE
1

                            Privacy in VoIP Networks:
                        Flow Analysis Attacks and Defense
                            Mudhakar Srivatsa† , Arun Iyengar† , Ling Liu‡ and Hongbo Jiang∗
                                           IBM T.J. Watson Research Center†
                                 College of Computing, Georgia Institute of Technology‡
                                     Huazhong Institute of Science and Technology∗
               {msrivats, aruni}@us.ibm.com, lingliu@cc.gatech.edu, hongbojiang@hust.edu.cn

    Abstract— 1 Peer-to-Peer VoIP (voice over IP) networks, ex-               in the network and to place voice calls to other clients on
emplified by Skype [5], are becoming increasingly popular due                 the network. VoIP uses the two main protocols: route setup
to their significant cost advantage and richer call forwarding                protocol for call setup and termination, and Real-time Trans-
features than traditional public switched telephone networks.
One of the most important features of a VoIP network is pri-                  port Protocol (RTP) for media delivery. In order to satisfy
vacy (for VoIP clients). Unfortunately, most peer-to-peer VoIP                QoS requirements, a common solution used in peer-to-peer
networks neither provide personalization nor guarantee a quan-                VoIP networks is to use a route setup protocol that sets up
tifiable privacy level. In this paper we propose novel flow analysis          the shortest route on the VoIP network from a caller src to
attacks that demonstrate the vulnerabilities of peer-to-peer VoIP             a receiver dst2 . RTP is used to carry voice traffic between
networks to privacy attacks. We then address two important
challenges in designing privacy-aware VoIP networks: Can we                   the caller and the receiver along an established bi-directional
provide personalized privacy guarantees for VoIP clients that                 voice circuit.
allow them to select privacy requirements on a per-call basis?                   In such VoIP networks, preserving the anonymity of caller-
How to design VoIP protocols to support customizable privacy                  receiver pairs becomes a challenging problem. In this paper we
guarantee? This paper proposes practical solutions to address                 focus on attacks that attempt to infer the receiver for a given
these challenges using a quantifiable k-anonymity metric and a
privacy-aware VoIP route setup and route maintenance protocols.               VoIP call using traffic analysis on the media delivery phase.
We present detailed experimental evaluation that demonstrates                 We make two important contributions. First, we show that
the performance and scalability of our protocol, while meeting                using the shortest route (as against a random route) for routing
customizable privacy guarantees.                                              voice flows makes the anonymizing network vulnerable to flow
                                                                              analysis attacks. Second, we develop practical techniques to
                                                                              achieve quantifiable and customizable k-anonymity on VoIP
                         I. I NTRODUCTION
                                                                              networks. Our proposal exploits the fact that audio codecs
   The concept of a mix [9] was introduced by Chaum in 1981.                  (such as G.729A without silence suppression3 ) generate sta-
Since then several authors have used mix as a network routing                 tistically identical packet streams that can be mixed without
element to construct anonymizing networks such as Onion                       leaking much information to an external observer (see Fig 2).
Routing [16], Tor [10], Tarzan [15], or Freedom [7]. Mix                         The following portions of this paper are organized as fol-
network provides good anonymity for high latency communi-                     lows. We present a reference model for a VoIP network fol-
cations by routing network traffic through a number of nodes                  lowed by flow analysis attacks in Section III. Section IV pro-
with random delay and random routes. However, emerging                        vides a more concrete definition of k-anonymity and describes
applications such as VoIP, SSH, online gaming, etc have addi-                 an efficient anonymity-aware route setup protocol (AARSP).
tional quality of service (QoS) requirements; for instance ITU                We sketch an implementation of our proposal and present ex-
(International Telecommunication Union) recommends up to                      perimental results that quantify the performance and scalability
250ms one-way latency for interactive voice communication;                    of AARSP in Section V. We present related work in Section
recent case study [26] indicates that latencies up to 250ms                   VI and finally conclude in Section VII.
are unperceivable to human users; while latencies over 400ms
significantly deteriorate the quality of voice conversations.
                                                                                              II. VO IP ROUTE S ETUP P ROTOCOL
   This paper examines anonymity for QoS sensitive appli-
cations on mix networks using peer-to-peer VoIP service as                       In this section, we describe a commonly used shortest route
a sample application. A peer-to-peer VoIP network typically                   setup protocol in peer-to-peer VoIP networks. The protocol
consists of a core proxy network and a set of clients that                    operates in four steps: initSearch (initiates a route setup by
connect to the edge of this proxy network (see Fig 1). This                   src), processSearch (process route setup request at some
network allows a client to dynamically connect to any proxy                     2 Enterprise VoIP networks that use SIP or H.323 signaling protocol may

   1 A short version of this paper appears in IEEE INFOCOM 2009:
                                                                              not use the shortest route
                                                                                3 G.729A without silence suppression deterministically generates one IP
http://www.research.ibm.com/people/i/iyengar/INFOCOM2009-kanon.pdf.           packet every 20ms. With silence suppression voice flows may be non-identical,
Prof. Ling Liu’s work was partially supported by grants from NSF CyberTrust   thereby making them trivially vulnerable to traditional traffic analysis attacks
program, AFOSR, Intel, and an IBM faculty award
2

Fig. 1.   Anonymizing VoIP Network

                    Flow 1: x pkts per sec            x pkts per sec    Flow 1: x pkts per sec              y pkts per sec

                                                         Flow ?                                                Flow 2
                                                         Flow ?                                                Flow 1

                    Flow 2: x pkts per sec            x pkts per sec   Flow 2: y pkts per sec               x pkts per sec
                                            VoIP Flows                       Unequal Flows: |x−y| > threshold
Fig. 2.   Mixing Statistically Identical VoIP Flows

node), processResult (process results of a route setup re-             from the following facts: (i) the first search request that reaches
quest at some node), and finSearch (concludes the route                a node p must have traveled along the shortest route from src
setup). One should note that flow analysis attacks exploits only       to p, and (ii) In processSearch a node p records the neigh-
the shortest path property and is independent of the concrete          bor q through which it received the first search request. This
route setup protocol. The description here serves as a basis for       indicates that the shortest route from src to p is via q. Setting
our AARSP in Section IV.                                               p = dst shows that route setup procedure in processResult
initSearch. A VoIP client src initiates a route setup for              builds the shortest VoIP network path from src to dst.
a receiver dst by broadcasting search(searchId, sipurl =                  After a successful route setup, the clients src and dst ex-
dst.sipurl, ts = curT ime) to all nodes p ∈ ngh(src), where            change an end-to-end media encryption key and switch to the
ngh(src) denotes the neighbors of node src in the VoIP net-            media delivery phase. The media delivery phase additional
work. Each VoIP client is identified by an URL (say, sip:bob@          uses hop-by-hop re-encryption using pair-wise shared keys
example.com). The search identifier searchId is a long ran-            between neighboring proxy nodes in the VoIP network. An
domly chosen unique identifier and ts denotes the time stamp           external observer tapping into the VoIP network may observe
at which the search request was initiated.                             hsrcIP , dstIP , srcP ort, dstP ort, EKp,q (EKsrc,dst (media))i,
processSearch. Let us suppose p receives search(searchId,              where Kp,q denotes a pair-wise shared symmetric key between
sipurl, ts) from its neighbor q. If p has seen searchId in             neighboring nodes p and q on the route, Ksrc,dst denotes the
the recent past then it drops the search request. Otherwise,           end-to-end encryption key and media denotes an encoding of
p checks if sipurl is the URL of a VoIP client connected               media bits. Further, an observer may observe packet sizes4 and
to p. If yes, p returns its IP address using result(searchId,          statistics on inter-packet arrival times. Re-encryption essen-
p) to q. p broadcasts search(searchId, sipurl, ts) to all p0           tially guarantees unlinkability between EKp,q (EKsrc,dst (media))
∈ ngh(p)−{q} and caches the search identifier hsearchId,               and EKq,r (EKsrc,dst (media)) by examining the contents of
sipurl, qi in its recently seen list. Note that p0 has no knowl-       the transmitted packet.
edge of where the search request is initiated.
processResult. Let us suppose p receives result(searchId,                                III. F LOW A NALYSIS ATTACKS
q) from q. Note that p has no knowledge as to where the search            In this section, we describe flow analysis attacks on VoIP
result was initiated. p looks up its cache of recently seen search     networks. These attacks exploit the shortest path nature of
queries to locate hsearchId, sipurl, previ. p adds a routing           the voice flows to identify pairs of callers and receivers on
entry hsipurl, qi and forwards result(searchId, p) to prev.            the VoIP network. Similar to other security models for VoIP
finSearch. When src receives result(searchId, q) from q,               networks, we assume that the physical network infrastructure
it adds a routing entry hdst, qi to its routing table.                 is owned by an untrusted third party (say, tier one/two network
   The route setup protocol establishes the shortest overlay
                                                                         4 Media packet sizes obtained from a common encoding algorithm such as
network route between src and dst. This observation follows
                                                                       G.729A are typically identical; if not packets can be padded with random bits
3

                                                                   Fig. 5.   Call Hold Time Data
Fig. 3.   Inferring Number of Flows from Flow Volume
                                                                             p1   1        1   p3              p1   1        1   p3
                                                                                      p5   1                            p5   1
                                                                                  1
                                                                             p2                p4              p2                p4
                                                                   Fig. 6.   Tracing Voice Calls     Fig. 7.   Shortest Path Tracing

                                                                   varying from 24ms − 150ms with a mean of 74ms and is
                                                                   within 20% error margin from real world latency measure-
                                                                   ments [17]. The average route (shortest path) latency between
                                                                   any two nodes in the network is 170ms, while the worst case
                                                                   route latency is 225ms. Our experiments over NS-2 use a
                                                                   bursty packet delay model wherein 20% of the packets incur
Fig. 4.   Call Volume Data                                         an additional delay of up to 44% of average one-way latency
                                                                   [17]. In practice, the total cost of framing, decoding and hop-
                                                                   by-hop re-encryption amounts to about 1.4ms per voice packet
service provider). Hence, the VoIP service must route voice        on commodity hardware. In our simulations, we adjust link
flows on the untrusted network in a way that preserves the         latencies to reflect the cost of routing VoIP packets.
identities of callers and receivers from the untrusted network.       We generate voice traffic based on call volume and call
We assume that the untrusted network service provider (ad-         hold time distribution obtained from a large enterprise with
versary) is aware of the VoIP network topology [28][10] and        973 subscribers (averaged over a month) (see Figures 4 and
the flow rates on all links in the VoIP network [19][7]. The       5). The call volume is specified in Erlangs [18]: if the mean
network service provider can obtain VoIP topology and flow         arrival rate of new calls is λ per unit time and the mean
information using traffic analysis (see Fig 3) or using various    call holding time (duration of voice session) is h, then the
measurement based approaches (such as expanding ring search        traffic in Erlangs A = λh; for example, if total phone use in a
on the network topology) [23]. We experimentally show that         given area per hour is 180 minutes, this represents 180/60 =
the attack can be very effective even when only one third of       3 Erlangs. We use G.729A audio codec for generating audio
the links are monitored by the adversary.                          traffic. The (src, dst) pair information for each call was not
   We represent the VoIP network topology as a weighted            made available. We have experimented under two settings:
graph G = hV , Ei, where V is the set of nodes and E ⊆             first, we assume that for a given VoIP call, the (src, dst) pair is
V ×V is the set of undirected edges. The weight of an edge e       chosen randomly from the VoIP network; second, we assume
= (p, q) (denoted by w(p, q)) is the latency between the nodes     that 80% of the calls are made between nodes that are in the
p and q. We assume that the adversary can observe the network      same network geography (e.g., same autonomous system). As
and thus knows nf (p → q) the number of voice flows between        noted in Section III-E, any prior information (such as, 80%
two nodes p and q on the VoIP network such that (p, q) ∈ E.        of call volume is limited to local network geography) can be
   To illustrate the effectiveness of our flow analysis attacks,   used by the adversary to further enhance the efficacy of flow
we use a synthetic network topology with 1024 nodes. The           analysis attacks. Finally, we note that all results reported in
topology is constructed using the GT-ITM topology gener-           this paper have been averaged over seven independent runs.
ator [35][1] and our experiments were performed on NS-2
[2][3]. GT-ITM models network geography and the small world        A. Naive Tracing Algorithm
phenomenon (power law graph with parameter γ=2.1) [14][22].          Let src be the caller. We use a Boolean variable f (p) ∈ {0,
The topology generator assigns node-to-node round trip times       1} to denote whether the node p is reachable from src using
4

           T RACE(Graph G=hV , Ei, Caller src)
           (1)    for each vertex v ∈ V                                                     S HORTEST PATH T RACING(Graph G=hV , Ei, Caller
           (2)        f [v] = 0; label[v] = false                                           src)
           (3)    end for                                                                   (1)   for each vertex v ∈ V
           (4)    f [src] = 1; label[src] = true                                            (2)       dist[v] = ∞; label[v] = false; prev[v] =
           (5)    while pick a vertex v labeled true                                              null
           (6)        label[v] = false                                                      (3)   end for
           (7)        for each node u such that (u, v) ∈ E                                  (4)   dist[src] = 0; label[src] = true
           (8)              if (f [u] = 0)                                                  (5)   while pick a labeled vertex v with minimum
           (9)                   f [u] = 1; label[u] = true                                       dist[v]
           (10)             end if                                                          (6)       label[v] = false
           (11)        end for                                                              (7)       for each node u such that (u, v) ∈ E
           (12) end while                                                                   (8)            if (dist[u] < dist[v] + w(u, v))
                                                                                            (9)                 dist[u] = dist[v] + w(u, v)
                                                                                            (10)                prev[u] = {v}; label[u] = true
Fig. 8.   Naive Tracing Algorithm                                                           (11)           end if
                                                                                            (12)           if (dist[u] = dist[v] + w(u, v))
                                                                                            (13)                prev[u] = prev[u] ∪ {v}
the measured flows on the VoIP network. One can determine                                   (14)           end if
                                                                                            (15)      end for
f (p) for all nodes p in O(|E|) time as follows. The base case                              (16) end while
of the recursion is f (src) = 1. For any node q, we set f (q)                               (17) G1 = hV 1 , E 1 i: V 1 = V , E 1 = (u → v) ∀u ∈
to one if there exists a node p such that (p, q) ∈ E ∧ f (p) =                                    prev[v], ∀v ∈ V
1 ∧ nf (p → q) > 0.
   Let us consider a sample topology shown in Figure 6. For            Fig. 9.            Shortest Path Tracing Algorithm
the sake of simplicity assume that each edge has unit latency.
The label on the edges in Figure 6 indicates the number of
voice flows. A trace starting from caller p1 will result in f (p1 )                               1
= f (p2 ) = f (p3 ) = f (p4 ) = f (p5 ) = 1. Filtering out the VoIP
proxy nodes (p5 ) and the caller (p1 ), the clients p2 , p3 and p4
could be potential destinations for a call emerging from p1 .                                0.1
   However, the tracing algorithm does not consider the short-
est path nature of voice routes. Considering the shortest path
                                                                        Precision

nature of voice paths leads us to conclude that p2 is not a                                 0.01
possible receiver for a call from p1 . If indeed p2 were the
receiver then the voice flow would have taken the shorter route
p1 → p2 (latency = 1), rather than the longer route p1 → p5 →                              0.001
p2 (latency = 2) as indicated by the flow information. Hence,
we now have only two possible receivers, namely, p3 and p4 .                                                          ’naive-trace’
                                                                                                              ’shortest-path-trace’
                                                                                          0.0001
B. Shortest Path Tracing Algorithm                                                                    1   2   4    8     16 32 64 128 256 512
                                                                                                                  Call Traffic (erlangs)
   In this section, we describe techniques to generate a directed
sub-graph G1 = hV 1 , E 1 i from G which encodes the shortest          Fig. 10.            Shortest Path Tracing Vs Naive Tracing
path nature of the voice paths. Given a graph G and a caller
src, we construct a sub-graph G1 that contains only those
voice paths that respect the shortest path property. Figure 9                                1
uses a breadth first search on G to compute G1 in O(|E|)                                                                              ’recall’
                                                                                                                                  ’precision’
time.                                                                                                                           ’f-measure’
   One can formally show that the directed graph G1 satisfies
the following properties: (i) If the voice traffic from src were
                                                                        Attack Efficacy

                                                                                            0.1
to traverse an edge e ∈   / E 1 , then it violates the shortest path
property. (ii) All voice paths that respect the shortest path
property are included in G1 . (iii) The graph G1 is acyclic.
   Figure 7 illustrates the result of applying the algorithm in                            0.01
Figure 9 on the sample topology in Figure 6. Indeed if one uses
the trace algorithm (Figure 8) on graph G1 , we get f (p2 ) =
0, f (p3 ) = f (p4 ) = 1. Figure 10 compares the effectiveness of
the shortest path tracing algorithm with the tracing algorithm                            0.001
                                                                                                  1                  2                      4
on a 1024 node VoIP network. On the x-axis we plot the
call traffic measured in Erlangs. We quantify the efficacy of                                                        Parameter k
an attack using standard metrics from inference algorithms:            Fig. 11.            Statistical Shortest Path Tracing: 128 Erlang Call Volume
precision, recall and F-measure. We use S to denote the set
5

                  1                                                      shortest paths from src to v. One can formally show that all
                                                                         voice paths that respect the top-κ shortest path property are
                                                                         included in Gκ . However, unlike G1 , the graph Gκ (for κ ≥
                0.1                                                      2) may contain directed cycles.
                                                                            Evidently, as κ increases, the tracing algorithm can accom-
 F-measure

                                                                         modate higher uncertainty in network latencies, thereby im-
               0.01                                                      proving recall. On the other hand, as κ increases, the precision
                                                                         initially increases and then decreases. The initial increase is
                                                                         attributed to the fact when κ is small the tracing algorithm may
              0.001                                                      even fail to identify the actual receiver as a candidate receiver;
                                                                         f [dst] may be 0 resulting in zero precision. However, for
                                            ’shortest-path-trace’
                                       ’statistical-shortest-path’       large values of κ, the number of candidate receivers becomes
             0.0001                                                      very large, thereby decreasing the precision metric. Figure 11
                      1      2     4    8     16 32 64 128 256 512
                                                                         shows the precision, recall and F -measure of the statistical
                                       Call Volume (Erlang)
                                                                         shortest path tracing algorithm with 128 Erlang call volume
Fig. 12.      F -measure with κ = 2                                      and varying κ. This experiment leads us to conclude that κ =
                                                                         2 yields a concise and yet precise list of potential receivers;
                                                                         observe that κ = 2 improves precision and recall by 97% and
of nodes such that for every p ∈ S, f [p] = 1. Recall denotes the        37.5% (respectively) over κ = 1.
probability of identifying the true receiver dst (dst ∈ S?), and            Figure 12 compares the F -measure for the statistical short-
precision is inversely related to the size of candidate receiver         est path tracing algorithm (κ = 2) and the shortest path tracing
        1                                                                algorithm (κ = 1) and varying call volume. We observe that
set (∝ |S| ). F -measure (computed as harmonic mean of recall
and precision scores) is a commonly used as a single metric              for low call volumes the shortest path tracing algorithm is
for measuring the effectiveness of inference algorithms [29].            sufficiently accurate. However, for moderate call volumes the
                                                                         statistical shortest path tracing algorithm can improve attack
                          recall
                          = P r(f [dst] = 1)                             efficacy by 1.5-2.5 times.
                            (
                               1
                              |S|   if dst ∈ S
                precision =
                              0     otherwise                            D. Flow Analysis Algorithm
                            2 ∗ recall ∗ precision                          We have so far used a Boolean variable f (p) to denote
                F-measure =
                             recall + precision                          whether a VoIP client p can be a potential receiver for a VoIP
In a deterministic network setting, the receiver dst is guar-            call from src. In this section, we use the flow measurements
anteed to be marked with f [dst] = 1, that is, recall = 1 for            to construct a probability distribution over the set of possible
both the naive tracking algorithm and the shortest path tracing          receivers. Let Gκ be a sub-graph of G obtained using the top-κ
algorithm. Hence, Figure 10 compares only the precision of               shortest path tracing algorithm with caller src. Let nf (p → q)
these two algorithms. We observe that for low call volumes (<            denote the number of flows on the edge p → q. Let in(p)
64 Erlangs) the shortest path tracing algorithm is about 5-10            denote the total number of flows into node p. Note that both
times more precise than the naive tracking algorithm. However,           nf (p → q) and in(p) are observable by an external adversary.
higher call volumes facilitate natural mixing of VoIP flows              Assuming a node p in the VoIP network performs perfect
thereby decreasing the precision of both the naive tracing and           mixing, the probability that some incoming flow is forwarded
shortest path tracing algorithms.                                        on the edge p → q as observed by an external adversary is
                                                                         nf (p→q)
                                                                           in(p) . Let f (p) denote the probability that a VoIP flow
                                                                         originating at src flows through node p. The function f is
C. Statistical Shortest Path Tracing                                     recursively defined on the directed edges in Gκ =hV κ , E κ i as
   In a realistic setting with uncertainties in network latencies        follows:
the shortest path tracing algorithm may not identify the re-
                                                                                                  X              nf (p → q)
                                                                                       f (q) =           f (p) ∗                      (1)
ceiver. We handle such uncertainties in network link latencies                                         κ
                                                                                                                    in(p)
                                                                                               p→q∈E
by using a top-κ shortest path algorithm to construct Gκ from
G. An edge e is in Gκ if and only if it appears in some top-κ            with the base case f (src) = 1 and in(src) = 1. Now, every
shortest path originating from src in graph G. We modified               VoIP client p (p 6= src) is a possible destination for the VoIP
Algorithm 9 to construct Gκ by simply maintaining top-κ                  flow originating from src if f (p) > 0. We use the top-m
distance measurements dist1 [v], dist2 [v], · · · , distκ [v] instead    probability metric, namely, the probability that the receiver
of only the shortest (top-1) distance measurement. We also               dst appears in the top-m entries when f (p) is sorted in de-
maintain previous hops prev 1 [v], prev 2 [v], · · · , prev κ [v] that   scending order. Top-m probability besides being an intuitive
correspond to each of these top-κ shortest paths. We add an              metric directly relates to the information theoretic notion of
edge (u, v) to E κ if u = prev i [v] for some 1 ≤ i ≤ κ. We              self-information (entropy) I = − log p. Indeed higher entropy
say that the voice traffic from src to v satisfies the top-κ             represents higher randomness and thus translates into better
shortest path property if it is routed along one of the top-κ            privacy. Such probability and entropy based metrics are often
                                                                         used for quantifying privacy in anonymous networks [19].
6

                       1                                                                                 1
                      0.9                                                                               0.9
                      0.8                                                                               0.8
 Top-10 Probability

                                                                                    Top-m Probability
                      0.7                                                                               0.7
                      0.6                                                                               0.6
                      0.5                                                                               0.5
                      0.4                                                                               0.4
                      0.3                                                                               0.3
                      0.2       ’spt’                                                                   0.2        ’m=1’
                                  ’fa’                                                                            ’m=10’
                      0.1       ’fah’                                                                   0.1       ’m=20’
                                ’fad’                                                                             ’m=50’
                       0                                                                                 0
                            1     2      4    8    16 32 64         128 256 512                               1     2       4    8    16 32 64         128 256 512
                                             Call Volume (Erlang)                                                               Call Volume (Erlang)

Fig. 13. Top-10 Probability (κ = 2): Comparison between shortest path              Fig. 14.              Top-m Probability with κ = 2 and Distance prior
tracing, flow analysis algorithm with no prior and Distance prior
                                                                                                                     κ      # Iterations   Time (ms)
                                                                                                                     1            -           0.03
    Computing the probabilities f (p) for G1 (top-1 shortest                                                         2           7             31
paths) is very efficient. Observe that since G1 is a directed                                                        3           11           49.7
acyclic graph, it can be sorted topologically. Let p1 = src, p2 ,                                                    4           19           84.2
· · · pN be a topological ordering of the nodes in G1 such that
                                                                                   Fig. 15.              Computation Cost
f (pi ) depends on f (pj ) only if j < i. Hence, one can effi-
ciently evaluate the probabilities by following the topological
order, namely, compute f (p2 ), f (p3 ), · · · , f (pN ) in that order                                   1
                                                                                                                   ’m=1’
starting with f (p1 ) = 1.                                                                              0.9       ’m=10’
    However, Gκ (for κ ≥ 2) may contain cycles and thus                                                           ’m=20’
                                                                                                        0.8       ’m=50’
cannot be topologically sorted. In this case, we represent the
                                                                                    Top-m Probability

                                                                                                        0.7
set of equations in 1 as π = πM , where π is a 1×N row
                                                                                                        0.6
vector and M is a N ×N matrix, where πi = f (pi ) and Mij
     nf (pi →pj )                                                                                       0.5
= in(p      i)
                  if there exists a directed edge pi → pj in Gκ ;
and Mij = 0 otherwise. Hence, the solution π is the stationary                                          0.4
distribution of a Markov chain whose transition probability                                             0.3
matrix is given by M . We compute π using a fixed point com-                                            0.2
putation approach as follows: we recursively compute π t+1 =                                            0.1
π t M starting with πi0 = 1 if pi = src; πi0 = 0 otherwise.                                              0
Assuming M is irreducible, π converges to a steady state                                                      0          0.2       0.4         0.6       0.8    1
solution π ∗ in O(N log N ) iterations. If the matrix M were                                                                Fraction of Flow Information
not irreducible (this happens if the underlying directed graph
G  κ
      is not strongly connected5 ), then we approximate π ∗ as
P τ ∗N log N t
                                                                                   Fig. 16.              Incomplete Flow Information
               π
    t=0
   τ ∗N log N     for some constant τ ≥ 1 (typically, we set τ =
1).                                                                                                      1

                                                                                                        0.9
E. Distance Prior and Hop Count Prior
                                                                                                        0.8
                                                                                    Top-m Probability

   In this section, we enhance the efficacy of the flow analysis
algorithm using hop count and distance prior. We use ghop                                               0.7
and glat to denote hop count and distance (in terms of latency)
between src and dst. For instance, one can use ghop and glat to                                         0.6
encode the fact that most calls are between nodes in the same
                                                                                                        0.5
autonomous system. Using the hop count prior, the probability                                                                                       ’m=1’
that a node p forwards an incoming flow on the edge p →                                                 0.4                                        ’m=10’
         (p→q)                                                                                                                                     ’m=20’
q is nfin(p)    ∗ P r(hop ≥ hc(src, p) + 1 | hop ≥ hc(src,                                                                                         ’m=50’
p)), where hc(src, p) denotes the number of hops along the                                              0.3
                                                                                                         0.0625                 0.125           0.25           0.5
shortest path between src and p on graph Gκ . P r(hop ≥
                                                                                                                           Fraction of Compromised Nodes
  5 A directed graph is said to be strongly connected if there exists a directed
path from every vertex to every other vertex in the graph                          Fig. 17.              Compromised Proxies
7

hc(src, p) + 1 | hop ≥ hc(src, p)) = PPr(hop≥hc(src,p)+1)
                                               r(hop≥hc(src,p))                  = hV κ , E κ i and the set of unmonitored links U ⊆ E κ as
denotes the probability that the receiver dst could be one more                  follows:
hop away given that dst is at least hc(src, p) hops away from                             X                1       X               nf (p → q)
src.                                                                             f (q) =         f (p) ∗        +          f (p) ∗
                                                                                                         deg(p)        κ
                                                                                                                                      in(p)
   A similar analysis applies to distance prior as well. We use                           p→q∈U                           p→q∈E \U

dist(src, p) to denote the latency of the shortest path between                  Figure 16 shows the efficacy of the attack assuming that only a
src and p on graph Gκ , and w(p, q) denotes the one-way                          fraction of the flow information is monitored by the adversary.
latency between nodes p and q. As with the flow analysis                         It follows from the figure that with only 20% flow information,
algorithm, the function f is defined on the directed edges in                    flow analysis attacks are ineffective; however, with 30-60%
graph Gκ = hV κ , E κ i as follows:                                              flow information the top-m probability increases substantially.
                                                                                 This suggests that the flow analysis attack is robust against
                X                 nf (p → q) P r(hop ≥ hc(src, p) + 1)
f (q)    =              f (p) ∗
                                    in(p)
                                            ∗
                                               P r(hop ≥ hc(src, p))
                                                                                 some inaccuracies in topology and flow information.
              p→q∈E κ
                X                 nf (p → q) P r(lat ≥ dist(src, p) + w(p, q))
f (q)    =              f (p) ∗             ∗
              p→q∈E κ
                                    in(p)         P r(lat ≥ dist(src, p))        G. Compromised Proxies
                                                                                    In addition, to passive observation based attacks, the ad-
   Figure 13 shows the Top-10 probability of an attack versus                    versary could actively compromise some of the nodes in the
call volume using κ = 2 for different attack algorithms spt:                     VoIP proxy. We assume an honest-but-curious model for the
statistical shortest path tracing, fa: flow analysis (with no                    compromised nodes. A compromised node p reveals its mixing
priors), fah: flow analysis with hop count prior and fad: flow                   information (namely, nf (r → p → q)) to an adversary, where
analysis with distance prior. We observe that the flow analysis                  nf (r → p → q) denotes the number of voice flows that were
algorithm with distance prior offers the best results. This is                   routed from r to q by the malicious node p. With slight abuse
primarily because the route setup protocol always constructs                     of notation, we use f (p → q) to denote the probability that a
voice paths that have minimum one-way latency. The distance                      VoIP call from src traverses the edge p → q. Hence, the new
prior directly reflects on the latency based shortest path nature                flow analysis equations are as follows:
of the route setup protocol and thus performs best.
   Figure 14 shows the top-m probability, namely, the proba-
                                                                                                     (             nf (p→q)
                                                                                                         f (p) ∗    in(p)
                                                                                                                                           honest p
bility that the true receiver dst appears in the top-m entries                    f (p → q)     =        P                    nf (r→p→q)
                                                                                                           r→p f (r → p) ∗     nf (r→p)
                                                                                                                                           malicious p
when f (p) is sorted in descending order using the flow anal-                                        X
ysis algorithm with distance prior and κ = 2. We also experi-                           f (q)   =          f (p → q)                                   (2)
                                                                                                     p→q
mented with hop count prior; however, distance prior directly
reflects on the latency based shortest path nature of the route                     Figure 17 shows the effectiveness of the attack as we vary
setup protocol and thus performs best. With a call volume of                     the fraction of nodes that is compromised by the adversary.
64 Erlangs, there is 86% chance that the true receiver dst                       We use a call volume of 128 Erlangs and κ = 2 in this
appears in the top-10 entries. Under very high call volume                       experiment. Evidently, compromised proxy nodes significantly
(512 Erlangs) the top-10 probability drops to 0.17. However,                     enhance attack efficacy; for instance, when 20% of the nodes
we note from our enterprise data set (see Fig 4) that the call                   are compromised the top-1 probability improves from 0.23 to
volume is smaller than 64 Erlangs for about 75% of the day.                      0.50.
   Figure 15 shows the computation cost6 incurred in com-
puting the probabilities f (p) for all potential receivers p. In
                                                                                           IV. VO IP P RIVACY USING k-A NONYMITY
practice, we observed that the number of iterations required
for f (p) to converge to its stationary value is much smaller                       In this section, we develop a k-anonymity approach to pro-
that the theoretical bound O(N log N ). We attribute this to the                 tect the identity of a receiver from flow analysis attacks. We
sparse nature of matrix M . The overall running time is of the                   define k-anonymity for identical voice flows as follows:
order of few tens of milliseconds making the attack feasible                     k-anonymity: A voice flow from src to dst is said to be k-
in real-time.                                                                    anonymous if the size of a candidate receiver set identified by
                                                                                 an adversary using the naive tracking algorithm is no smaller
F. Incomplete Flow Information                                                   than k 7 .
   So far, we have assumed that the adversary can monitor the                    As pointed out earlier, k-anonymity can be realized by mixing
flow rate on all the links in the VoIP network. In the absence                   a voice flow from src to dst with at least k − 1 other flows
of flow information on a link p→q, we use an unbiased ran-                       (each of which has a different source and destination nodes).
dom walk estimator to compute f (p)’s contribution to f (q) as                   The primary supposition that any two voice flows are indistin-
           1
f (p) ∗ deg(p) , where deg(p) denotes the number of neighbors                    guishable under traffic analysis (see Fig 2) follows from the
of p on the VoIP network. As with the flow analysis algorithm,                   properties of RTP (real-time protocol) and the constant packet
the function f is defined on the directed edges in graph Gκ                      rate property of silence suppression free audio codecs. One
  6 As   measured on a 900 MHz Pentium III processor running RedHat Linux          7 k-anonymity is defined with respect to naive tracing, since AARSP does
7.2                                                                              not use shortest path routing
8

way to achieve k-anonymity is to mix a flow from src to dst                                                      p3 0
with k − 1 dummy voice flows; however, this approach can                                                   0
increase aggregation bandwidth consumption by k-fold. In this
section, we propose an anonymity aware route set up protocol                                        p1                     p2
(AARSP). AARSP reroutes and mixes existing voice flows                                                     1                      1
                                                                                            1                         1
(without adding dummy traffic) with the goal of: (i) meeting
k-anonymity, and (ii) satisfying latency based QoS guarantee.
   Figures 18 illustrates a simple scenario with two VoIP flows                              s1            d1         s2          d2
between clients (s1 , d1 ) and (s2 , d2 ). In the figures, each edge
                                                                                            Fig. 18.   1-anonymity
on the VoIP network is marked with the number of flows it
carries. Figure 19 shows how one can reroute VoIP traffic in
order to achieve 2-anonymity. Evidently, the rerouted flows                                                1    p3   1
may not satisfy the shortest route property and thus may vio-
late latency based QoS guarantees. However, AARSP exploits
the slack between the shortest path latency and the tolerable
                                                                                                    p1                     p2
latency (250ms) to accomplish its goals.                                                    1              1         1             1
A. AARSP: Anonymity-Aware RSP
   In this section, we summarize our ideas behind AARSP.
                                                                                            s1             d1        s2           d2
AARSP accepts an anonymity parameter k as an input for the                                  Fig. 19.   2-anonymity
route setup protocol, on a per-client per-call basis. AARSP
modifies the basic route set up protocol (RSP) such that it si-
multaneously satisfies three conditions: (i) we have at least one      respectively. Node p drops the search request received at time
node p ∈ route(src, dst) such that in(p) ≥ k (k-anonymity),            t2 if: prev1 = prev2 .
(ii) the end-to-end one-way latency on the route from src to              AARSP sets up the shortest possible route from src to dst
dst is smaller than maxLat (typically set to 250ms), and (iii)         that meets k-anonymity (if there exists such a route). If no
the total call volume on every node p ∈ route(src, dst) is             such route exists, AARSP setup identifies a route with highest
smaller than its capacity maxF low(p).                                 possible anonymity k 0 < k; in addition, AARSP sets up the
   Similar to the naive route set up protocol discussed in Sec-        shortest route that achieves k 0 -anonymous route from src to
tion II, AARSP uses an expanding search approach to con-               dst. Indeed, if one sets k = ∞, then AARSP identifies the
struct a k-anonymous path from a caller src to a recipient             most anonymous route whose one-way latency is smaller than
dst. The key idea is that the expanding search not only tracks         maxLat = 250ms. For detailed analysis on the properties
the distance from src to a node p in the network (dist[p]), but        of the AARSP protocol we refer the readers to a detailed
also the anonymity of a route from src to p (anon[p]). Recall          technical report [27].
that the naive route set up protocol drops a search request if            Figure 20 compares the average one-way path latency for
it has seen the search identifier searchId in the recent past.         both AARSP and RSP for varying call volumes. At low call
On the other hand, in the AARSP protocol, let us suppose               volumes AARSP may construct significantly longer routes than
that a node p receives a search request identified by searchId         RSP. At lower call volumes, AARSP may significantly de-
at two time instants t1 and t2 (wlog, t1 < t2 ) such that the          viate from the shortest path to achieve the desired level of
anonymity of the route traversed by the search requests is k1
and k2 respectively. Node p drops the search request received
at time t2 if: k1 ≥ k or k1 ≥ k2 .                                                250
                                                                                                                                 ’rsp’
   AARSP may intentionally introduce loops in the path to                         240                                         ’aarsp’
improve the anonymity of voice flows. In doing so, AARSP                          230
ensures the protocol converges to a valid path by limiting                        220
number of times any loop is traversed by a path to at most one.
                                                                                  210
                                                                        Latency

Unlike the shortest path route setup protocol, the routing entry
                                                                                  200
at a node p in the AARSP protocol is a three tuple hsipurl,
prev, nexti, which indicates that when a node p receives voice                    190
packets from node prev destined to sipurl, then node p must                       180
forward the packet to node next. The three tuple routing table                    170
allows us to handle loops; for example, from Figure 19 node                       160
p3 may have the following two routing entries: hd1 , p1 , p2 i
                                                                                  150
and hd1 , p2 , p1 i. In the AARSP protocol, let us suppose that                         1       2      4    8    16 32 64         128 256 512
a node p receives a search request identified by searchId at                                               Call Volume (Erlang)
two time instants t1 and t2 (wlog, t1 < t2 ) such that the search
request was forwarded to node p by nodes prev1 and prev2               Fig. 20.     AARSP Vs RSP: Path Latency
9

anonymity; higher call volumes facilitate natural mixing of
voice flows and thus require relatively smaller deviation from                            30
the shortest path. Nonetheless, AARSP ensures that the la-
tency is below 250ms and thus preserves the quality of voice                              25
conversations.

                                                                     Anonymity Level
   Figure 21 shows the anonymity level of a VoIP call as                                  20
a function of its hold time given that the initial route was
established using k = 20. Anonymity level of a VoIP flow F                                15
at time t denotes the number of flows mixed with F at time
t along F ’s route. We tracked the worst case (lowest) and                                10
the best case (highest) anonymity levels of a VoIP call. We
                                                                                                            ’anon-low-64Erlang’
observe that at lower call volumes (64 Erlangs) calls with hold                           5               ’anon-low-128Erlang’
time larger than 400 seconds may experience time durations                                                 ’anon-high-64Erlang’
                                                                                                         ’anon-high-128Erlang’
at which their anonymity level falls below 80% of the initial                             0
value. Fortunately, 95% of voice calls are shorter than 400s                                           100                                  1000
(see Fig 5). Long lasting calls may set up a lower watermark                                                          Call Hold Time (seconds)
level k 0 ≤ k such that the AARSP route maintenance protocol
                                                                    Fig. 21.                   Route Maintenance
automatically terminates the call when its anonymity level falls
below k 0 .

B. Flow Analysis Attacks on AARSP
   A flow analysis attack on AARSP operates efficiently as                                     1
follows. We assume that the adversary knows the parameter k                               0.9
used by a caller src. First, the adversary identifies all nodes p
                                                                                          0.8
such that in(p) ≥ k; if there exists no such p, then the attacker
                                                                     Top-10 Probability

picks p that maximizes in(p). Let Gκsrc be a sub-graph of G                               0.7
obtained using the top-κ shortest path tracing algorithm with                             0.6
caller src. We use Gκsrc and the flow measurements to compute                             0.5
the probability that the voice flow is routed from src to p
(fsrc (p)) for all p such that in(p) ≥ k starting with fsrc (src)                         0.4
= 1. The algorithm used for computing fsrc (p) is described                               0.3
in Section III-D.                                                                         0.2             ’rsp’
   Second, for every such node p with in(p) ≥ k, we con-                                               ’aarsp’
                                                                                          0.1
struct Gκp as a sub-graph of G obtained using the top-κ short-                                     1    2         4      8    16 32 64         128 256 512
est path tracing algorithm with p as the caller. We compute                                                             Call Volume (Erlang)
the probability of a voice flow being routed from src to r
via p (fsrc,p (r)) for every candidate receiver r starting with     Fig. 22.                   AARSP Vs RSP: Top-10 Probability
fsrc,p (p) = fsrc (p), where fsrc (p) is obtained from the first
step. Third, we compute the probability of r being the re-
ceiver as frecv (r) = fsrc,pivot (r), where pivot = argminp
{dist(src, p) + dist(p, r)}. Similar to other flow analysis at-
                                                                                               1
tacks, one could sort the receivers in descending order of f (r)                                          ’rsp’
                                                                                          0.9          ’aarsp’
and use a top-m probability metric to study the efficacy of the
attack.                                                                                   0.8
                                                                     Top-m Probability

   Figures 22 and 23 compare the effectiveness of AARSP                                   0.7
with the shortest path route setup protocol (RSP) in mitigating                           0.6
flow analysis attacks. We set the parameter k = ∞ so that                                 0.5
AARSP identifies the most anonymous route from caller src
                                                                                          0.4
to receiver dst that satisfies the latency constraint. We use the
flow analysis described in this section for AARSP and a flow                              0.3
analysis attack with distance prior for RSP. Figure 22 shows                              0.2
the top-10 probability using both AARSP and RSP for varying                               0.1
call volumes. Observe that at low and moderate call volumes                                    0
AARSP offers significantly improved protection against flow                                        1         2           4        8      16        32   64
analysis attacks. Figure 23 shows the top-m probability for                                                                  Parameter m
both AARSP and RSP for varying m and a call volume of
                                                                    Fig. 23.                   AARSP Vs RSP: 128 Erlangs
128 Erlangs. We also observe that AARSP consistently out
performs RSP for all values of m.
10

C. Tolerating Compromised Proxies                                                               1
    In this section, we study the effect of compromised proxies                  0.9
on AARSP. Similar to Section III, we assume that a com-
                                                                                 0.8
promised node will execute AARSP honestly. However, the

                                                                           Top-10 Probability
malicious nodes may record observed information during the                       0.7
voice session and collude with the external observer. In par-                    0.6
ticular, a malicious node may reveal the mapping between its                     0.5
in-flows and out-flows thereby decreasing the anonymity of a
VoIP flow.                                                                       0.4
    We describe a simple extension to AARSP to tolerate ma-                      0.3
                                                                                                                              ’c=1’
licious nodes and yet preserve k-anonymity. We allow the                         0.2                                          ’c=2’
caller src to specify a personalized security parameter c for                                                                 ’c=3’
                                                                                 0.1
every VoIP call indicating that the route from src to dst                         0.0625               0.125             0.25              0.5
should tolerate up to c compromised nodes while preserving                                        Fraction of Compromised Nodes
k-anonymity. The key idea is to construct a route from src to
dst with at least c + 1 nodes p1 , p2 , · · · , pc+1 such that in(pi ) Fig. 24. Compromised Proxies
≥ k for all 1 ≤ i ≤ c+1. One can introduce this constraint by
recording top-c anonymity levels in the anon field of ASRSP.                                            k Time(s)
Now, one could replace the constraint anon ≥ k with anon                                                2        9.0
≥ (k, c).                                                                                              10        5.9
    While this approach offers significantly higher attack re-                                         20        4.0
silience, one can extend the flow analysis attack on a k-anonymous                                     50        2.9
AARSP to a (k, c)-anonymous AARSP. We assume that the Fig. 25. Attack Cost Vs Anonymity Level (k)
adversary knows the parameters k and c used by a caller
src. The adversary identifies a set of nodes S such that for
all nodes p ∈ S, in(p) ≥ k. For every permutation of c computation cost for an attack grows rapidly with c, making
nodes from the set S, the adversary computes the probabil- it harder for an adversary to launch flow analysis attacks in
ity of a voice flow being routed from src to r via p1 , p2 , real-time.
· · · , pc (fsrc,p1 ,p2 ,··· ,pc (r)) for a candidate receiver r. We
recursively compute fsrc,p1 ,p2 ,··· ,pi (r) using flow analysis al-                       V. P ERFORMANCE E VALUATION
gorithms on Gκpi and starting probability fsrc,p1 ,··· ,pi (pi ) = A. Implementation Sketch
fsrc,p1 ,p2 ,··· ,pi−1 (pi ). The recursion terminates at the base case
fsrc (src) = 1. Finally, we compute the probability of r be-               In this section, we briefly describe an implementation of
ing the receiver as frecv (r) = fsrc,pivot (r), where pivot =           our   algorithms using Phex [4]: an open source Java based
argmin{p1 ,p2 ,··· ,pc }⊆S {latdist(src, p1 ) + latdist(p1 , p2 ) +     implementation        of shortest route set up protocol (RSP). VoIP
· · · + latdist(pc−1 , pc ) + latdist(pc , r)}. This attack reduces     protocols     operate    on top of the peer-to-peer infrastructure.
a simple flow analysis attack on AARSP when c = 1; however,             We   have    implemented       our algorithms as pluggable modules
the pivot size and consequently the attack complexity grows             that  can   be   weaved    into  the Phex client code using AspectJ
exponentially in c.                                                     [12].  Our   implementation      is completely transparent to the VoIP
    Figure 24 shows the effectiveness of the attack as we vary          protocol    that  operates  on  top  of the peer-to-peer infrastructure.
the fraction of nodes that are compromised by the adversary             Also,    our   implementation       does  not require any changes to
and the parameter c. We use a call volume of 128 Erlangs,               topology     construction    and    maintenance   algorithms (as nodes
k = 10 and κ = 2 in this experiment. We observe that as c               join,  leave,   fail or recover)   and the  underlying  TCP/IP or UDP
increases the effectiveness of the attack decreases significantly;      based    communication       libraries.
for instance, when 10% of the nodes are compromised, using                 Below we sketch our implementation of AARSP. The Phex
c = 3 reduces the top-10 probability to 0.06 when compared to           broadcast     search protocol has four operations: initSearch,
                                                                        processSearch, processResult and finSearch. These
c = 1 which results in a top-10 probability of 0.44. Nonethe-
less, when a large number of nodes are malicious AARSP four operations are implemented as event handlers in Phex.
becomes vulnerable to flow analysis attacks. Under a realistic When a Phex client receives a messages, it determines the
setting, when only a small number of nodes (< 20%) are likely type of the message (search request, search result, etc) and
to be malicious, AARSP offers very good protection against
flow analysis attacks.                                                                                  c Time(s)
    Figure 25 shows the computation cost incurred to an adver-                                          1       2.9
sary as k increases. As k increases, the number of pivot nodes                                          2       4.4
p, that is, p such that in(p) ≥ k decreases; and thus, the                                              3      11.6
computation cost decreases. Figure 26 shows the computation                                             4      43.5
cost for an attack as c increases with k=∞. We observe that the
                                                                         Fig. 26.               Attack Cost Vs Tolerance to Malicious Nodes (c)
11

                                                                                                         triggers the appropriate event handler. AARSP substitutes all
                                                                                                         four event handlers (see Section IV). In addition, we also mod-
 Per Node Msg Cost (# msgs per search)

                                                               ’rsp’                                     ify the search request payload and the search result payload
                                         2.5           ’aarsp-k=10’
                                                       ’aarsp-k=20’                                      to include the parameters k and c. In the rest of this section,
                                                       ’aarsp-k=50’                                      all experimental results are reported using our implementation
                                              2                                                          deployed on 64 nodes in Planet Lab8 .

                                         1.5
                                                                                                         B. Performance and Scalability
                                              1                                                             In this section, we compare the messaging cost and the
                                                                                                         search latency of AARSP and RSP. Figure 27 shows the mes-
                                         0.5
                                                                                                         saging cost per node as we vary the call volume and the
                                                                                                         anonymity parameter k (using c = 1). These experiments show
                                                                                                         that AARSP incurs about 1-3 times the messaging cost of RSP.
                                              0
                                                  1    2       4     8    16 32 64         128 256 512   However, the search request and the results are typically of
                                                                    Call Volume (Erlang)                 the order of 300 Bytes. Hence, the communication cost at a
                                                                                                         node for handling 10 messages (3 KB) is equivalent to a voice
Fig. 27.                                      Messaging Cost                                             session of one second (24 Kbps). Even though AARSP incurs
                                                                                                         higher communication cost than RSP, its effect on the overall
                                                                                                         bandwidth consumption is negligible.
                                                                                                            Figure 28 shows the latency of a search operation as we
                                         250                                                             vary the call volume and the anonymity parameter k (using c
                                                                                                         = 1). This experiment shows that AARSPs incurs about 30-
                                                                                                         40% higher search latency than RSP. One should note that the
                                         200                                                             search latency only affects the initial connection set up time.
 Search Latency (ms)

                                                                                                         Once the route is established AARSP ensures good quality
                                         150                                                             voice conversations by limiting the path latency to 250ms.
                                                                                                         Figures 27 and 28 also show that the relative overhead of
                                         100
                                                                                                         AARSP over RSP decreases with call volume. The key intu-
                                                                                                         ition here is that higher call volumes facilitate natural mixing
                                                                                                         of voice flows, thereby decreasing the overall messaging and
                                          50                   ’rsp’
                                                       ’aarsp-k=10’                                      search cost.
                                                       ’aarsp-k=20’                                         Figure 29 shows the average number of concurrent VoIP
                                                       ’aarsp-k=50’
                                              0                                                          calls handled by a node in the VoIP network (using c = 1).
                                                  1    2       4     8    16 32 64         128 256 512   AARSP incurs a higher load primarily because the traffic is
                                                                    Call Volume (Erlang)                 not routed through the optimal route. Hence, a voice route
                                                                                                         in AARSP may include more network hops than RSP. This
Fig. 28.                                      Search Latency
                                                                                                         increases the average number of proxies that route one voice
                                                                                                         call, and thus increases the average load on a proxy. The per-
                                                                                                         centile increase in average node load for higher call volumes
                                                                                                         is small. Hence, when the call volume is high (VoIP network
                                         45                                                              is heavily loaded), AARSP imposes small overheads (20-40%)
                                                              ’rsp’
                                         40           ’aarsp-k=10’                                       compared with RSP. On the other hand, when the call volume
 Node Load (# concurrent calls)

                                                      ’aarsp-k=20’                                       is low (VoIP network is lightly loaded), AARSP incurs 2-3
                                         35           ’aarsp-k=50’
                                                                                                         times the average node load when compared to RSP. However,
                                         30                                                              this 2-3x increase in node load is incurred when the VoIP
                                         25                                                              network is itself lightly loaded; hence, AARSP does not harm
                                                                                                         the performance and scalability of the VoIP network.
                                         20
                                         15
                                                                                                                                  VI. R ELATED W ORK
                                         10                                                                 Privacy has long been a hot button issue for both the VoIP
                                         5                                                               clients and the law enforcement bodies. On one hand, users
                                         0                                                               want their phone conversations to be anonymous; anonymity
                                              1       2     4       8     16   32    64    128 256 512   offers them possible deniability thereby shielding them from
                                                                   Call Volume (Erlangs)                 law enforcement bodies. On the other hand FCC (Federal
Fig. 29.                                      Node Load                                                    8 http://www.planet-lab.org/
12

Communications Commission) [13] considers capability of track- an anonymity aware route setup protocol (AARSP) that allows
ing VoIP calls “of paramount importance to the law enforce- clients to specify personalized privacy requirements for their
ment and the national security interests of the United States”. voice calls (on a per-client per-call basis) using a quantifiable
Similar to other VoIP privacy papers [30], we leave aside the k-anonymity metric. We have implemented our proposal on the
controversy between anonymity and security. Instead we focus Phex client and presented detailed experimental evaluation that
on technical feasibility of privacy attacks and defenses on VoIP demonstrates the performance and scalability of our protocol,
networks.                                                        while meeting customizable privacy guarantees.
   Mix [9] is a routing element that attempts to hide correspon-
dences between its input and output messages. A large number                                     R EFERENCES
of low latency anonymizing networks have been built using         [1] GT-ITM:        Georgia      tech   internetwork     topology    models.
the concept of a mix network [9][25]. Onion routing [16] and          http://www.cc.gatech.edu/projects/gtitm/.
its second generation Tor [10] aim at providing anonymous         [2] The network simulator NS-2. http://www.isi.edu/nsnam/ns/.
                                                                  [3] The       network       simulator    NS-2:      Topology     generation.
transport of TCP flows over the Internet. ISDN mixes [20]             http://www.isi.edu/nsnam/ns/ns-topogen.html.
proposes solutions to anonymize phone calls over traditional      [4] Phex client. http://www.phex.org.
PSTN (Public Switched Telephone Networks). In this paper we       [5] Skype - the global internet telephone company. http://www.skype.com.
                                                                  [6] Telegeography research. http://www.telegeography.com.
have focused on VoIP networks given its recent wide spread        [7] A. Back, I. Goldberg, and A. Shostack. Freedom 2.1 security issues and
adoption9 .                                                           analysis. Zero Knowledge Systems, Inc. white paper, 2001.
   It is widely acknowledged that low latency anonymizing         [8] A. Blum, D. Song, and S. Venkataraman. Detectiong of interactive
                                                                      stepping stones: Algorithms and confidence bounds. In 7th RAID, 2004.
networks [10][16][7] are vulnerable to timing analysis attacks    [9] D. Chaum. Untraceable electronic mail, return addresses, and digital
[28][25], especially from well placed malicious attackers [33].       pseudonyms. In Communications of ACM, 24(2): 84-88, 1981.
Several papers have addressed the problem of tracing encrypted [10] R. Dingledine, N. Mathewson, and P. Syverson. Tor: The second
traffic using timing analysis [32][34][36][31][8][11][30]. All [11] generation     onion router. In 13th USENIX Security Symposium, 2000.
                                                                      D. L. Donoho, A. G. Flesia, U. Shankar, V. Paxson, J. Coit, and
these papers use inter-packet timing characteristics for trac-        S. Staniford. Multiscale stepping stone detection: Detecting pairs of
ing traffic. Complementary to all these approaches, we have           jittered interactive streams by exploiting maximum tolerable delay. In
introduced flow analysis attacks that target the shortest path [12] 5th    RAID, 2002.
                                                                      Eclipse. Aspectj compiler. http://eclipse.org/aspectj.
property of voice routes and presented techniques to provide [13] FBI. Letter to FCC. http://www.askcalea.com/docs/20040128.jper.letter.pdf.
customizable anonymity guarantees in a VoIP network. Unlike [14] B. Fortz and M. Thorup. Optimizing OSPF/IS-IS weights in a changing
                                                                      world. In IEEE Journal on Special Areas in Communication, 2002.
the timing analysis attacks, our approach does not rely upon [15] M.       J. Freedman and R. Morris. Tarzan: A peer-to-peer anonymizing
inter-packet times to detect caller-receiver pairs; instead we        network layer. In 9th ACM CCS, 2002.
analyze the volume of flow in the VoIP network and deduce [16] D. Goldschlag, M. Reed, and P. Syverson. Onion routing for anonymous
                                                                      and private internet connections. In Communications of ACM, Vol 42(2),
possible caller-receiver pairs using the flow information and         1999.
the underlying VoIP network topology.                            [17] K. Gummadi, R. Gummadi, S. Gribble, S. Ratnasamy, S. Shenker, and
   Tarzan [15] presents an anonymizing network layer using a          I. Stoica. The impact of DHT routing geometry on resilience and
                                                                      proximity. In ACM SIGCOMM, 2003.
gossip-based peer-to-peer protocol. We note that flow analysis [18] J.    Y. Hui. Switching and traffic theory for integrated broadband
attacks target the shortest path property and not the protocol        networks. Academic Press ISBN: 079239061X, 1990.
used for constructing the route itself; hence, a gossip based [19] G. Perng, M. K. Reiter, and C. Wang. M2: Multicasting mixes for
                                                                      efficient and anonymous communication. In IEEE ICDCS, 2006.
shortest path setup protocol is equally vulnerable to flow anal- [20] A.  Pfitzmann, B. Pfitzmann, and M. Waidner. ISDN-MIXes: Untraceable
ysis attacks.                                                         communication with small bandwidth overhead. In GI/ITG Conference
   Traditionally, multicast and broadcast protocols have been         on Communication in Distributed Systems, 1991.
used to protect receiver anonymity [21][24]. However, in a [21] A.        Pfitzmann and M. Waidner. Networks without user observability. In
                                                                      Computers and Security, 2(6): 158-166.
multicast based approach achieving k-anonymity may increase [22] L. Qiu, V. N. Padmanabhan, and G. M. Voelker. On the placement of
the network traffic by k-fold. In contrast our paper attempts         web server replicas. In 12th IEEE INFOCOM, 2001.
to reroute and mix existing voice flows and thus incurs sig- [23] S.      Saroiu, P. K. Gummadi, and S. D. Gribble. A measurement study
                                                                      of peer-to-peer file sharing systems. In Multimedia Computing and
nificantly smaller overhead on the VoIP network.                      Networks (MMCN), 2002.
                                                                         [24] C. Shields and B. N. Levine. A protocol for anonymous communication
                                                                              over the internet. In ACM CCS, 2000.
                       VII. C ONCLUSION                                  [25] V. Shmatikov and M. H. Wang. Timing analysis in low latency mix
   In this paper we have addressed the problem of providing                   networks: Attacks and defenses. In 11th ESORICS, 2006.
                                                                         [26] G.      I.    Sound.               VoIP:    Better    than     PSTN?
privacy guarantees in peer-to-peer VoIP networks. First, we                   http://www.globalipsound.com/demo/tutorial.php.
have developed flow analysis attacks that allow an adversary             [27] M. Srivatsa, A. Iyengar, and L. Liu. Privacy in voip networks: A k-
(external observer) to identify a small and accurate set of                   anonymity approach. Technical report, IBM Research RC24625, 2008.
                                                                         [28] P. Syverson, G. Tsudik, M. Reed, and C. Landwehr. Towards an analysis
candidate receivers even when all the nodes in the network                    of onion routing security. In Workshop on Design Issues in Anonymity
are honest. We have used network flow analysis and statistical                and Unobservability, 2000.
inference to study the efficacy of such an attack. Second, we            [29] C. J. Van-Rijsbergen. Information retrieval. In Second Edition,
                                                                              Butterworths, London, 1979.
have developed mixing based techniques to provide a guaran-              [30] X. Wang, S. Chen, and S. Jajodia. Tracking anonymous peer-to-peer
teed level of anonymity for VoIP clients. We have developed                   VoIP calls on the internet. In 12th ACM CCS, 2005.
                                                                         [31] X. Wang and D. Reeves. Robust correlation of encrypted attack traffic
  9 According to TeleGeography Research [6], worldwide VoIP’s share of        through stepping stones by manipulation of interpacket delays. In 10th
voice traffic has grown from 12.8% in 2003 to about 44% in 2007               ACM CCS, 2003.
13

[32] X. Wang, D. Reeves, and S. Wu. Inter-packet delay based correlation for
     tracing encrypted connections through stepping stones. In 7th ESORICS,
     2002.
[33] X. Y. Wang and D. S. Reeves. Robust correlation of encrypted attack
     traffic through stepping stones by manipulation of interpacket delays. In
     ACM CCS, 2003.
[34] K. yoda and H. Etoh. Finding a connection chain for tracing intruders.
     In 6th ESORICS, 2000.
[35] E. W. Zegura, K. Calvert, and S. Bhattacharjee. How to model an
     internetwork? In IEEE Infocom, 1996.
[36] Y. Zhang and V. Paxon. Detecting stepping stones. In 9th USENIX
     Security Symposium, 2000.

                         Mudhakar Srivatsa is a research scientist at IBM
                         T. J. Watson Research Center where he conducts re-
                         search at the intersection of networking and security.
                         He received a B.Tech degree in Computer Science
       PLACE             and Engineering from IIT, Madras, and a PhD de-
       PHOTO             gree in Computer Science from Georgia Tech. His
        HERE             research interests primarily include security and re-
                         liability of large scale networks, secure information
                         flow, and risk management.

                         Arun Iyengar received the Ph.D. degree in com-
                         puter science from MIT. He does research and de-
                         velopment into Web performance, distributed com-
                         puting, and high availability at IBM’s T.J. Watson
       PLACE             Research Center. Arun is Co-Editor-in-Chief of the
       PHOTO             ACM Transactions on the Web, Founding Chair of
        HERE             IFIP Working Group 6.4 on Internet Applications
                         Engineering, and an IBM Master Inventor.

                           Ling Liu is an associate professor at the College
                           of Computing at the Georgia Institute of Technol-
                           ogy. There, she directs the research programs in the
                           Distributed Data Intensive Systems Lab (DiSL), ex-
        PLACE              amining research issues and technical challenges in
        PHOTO              building large scale distributed computing systems
         HERE              that can grow without limits. Her research group
                           has produced a number of software systems that
                           are either open sources or directly accessible on-
                           line, among which the most popular are WebCQ and
                           XWRAPElite. She is on the editorial board of sev-
eral international journals, such as the IEEE Transactions on Knowledge and
Data Engineering, the International Journal of Very large Database Systems
(VLDBJ), and the International Journal of Web Services Research.

                         Hongbo Jiang is an associate professor at the De-
                         partment of Electronics and Information Engineering
                         at the Huazhong University of Science and Technol-
                         ogy. Hongbo leads Netwoked and Communication
       PLACE             Systems Research group, wherein he conducts re-
       PHOTO             search on computer networking, especially focused
        HERE             on algorithms and architectures for high performance
                         wireless networks.
You can also read